摘要
针对现有的行人多目标跟踪算法在遮挡、人群密集和光线差等情况下表现不佳的问题,提出一种改进YOLOv4与改进DeepSort算法相结合的行人多目标跟踪算法。首先,为增强检测网络的特征提取能力,在YOLOv4中嵌入ECANet注意力模块,提高检测精度;其次,在改进DeepSort的跟踪算法中,由卡尔曼滤波算法预测多个行人目标在图像中的轨迹之后,使用GhostNetV1替换DeepSort中的重识别网络来生成行人的外观特征,提高行人重识别网络的性能;进而,采用匈牙利算法对检测框和预测框进行最优匹配,对未匹配成功的检测框采用DIOU代替IOU(交并比)进行二次匹配,提高DeepSort网络的跟踪性能;最后,开展了新跟踪算法与原DeepSort算法的对比实验,结果表明新算法的误检、漏检现象变少,鲁棒性增强,跟踪性能得到提高,MOTA提升了18.8%,IDF1提升了18.2%,身份编号转换次数降低了84次。
In allusion to the poor performance of existing pedestrian multi-target tracking algorithms in the case of occlusion,dense crowd and poor light,a pedestrian multi-target tracking algorithm based on improved YOLOv4 and improved DeepSort algorithms is proposed. In order to enhance the feature extraction ability of the detection network,ECANet attention module is embedded into YOLOv4 to improve the detection accuracy. In the improved tracking algorithm of DeepSort,the Kalman filter algorithm is used to predict the trajectories of multiple pedestrian targets in the image,and then GhostNetV1 is used to replace the re-recognition network in DeepSort to generate the appearance features of pedestrians,so as to improve the performance of the re-recognition network. The Hungarian algorithm is used to optimize the matching of detection boxes and prediction boxes,and DIOU(distance IOU loss)is used to replace IOU(overlap and union ratio)for the second matching of unmatched detection boxes,so as to improve the tracking performance of DeepSort network. The contrast experiment between the new tracking algorithm and the original DeepSort algorithm was carried out. The results show that the new tracking algorithm has less false detection and missing detection,its robustness is enhanced,its tracking performance is improved,its MOTA is increased by18.8% its IDF1 is increased by 18.2%,and the number of identity number conversion is decreased by 84.
作者
郑繁亭
邢关生
ZHENG Fanting;XING Guansheng(College of Automation and Electronic Engineering,Qingdao University of Science and Technology,Qingdao 266061,China)
出处
《现代电子技术》
2023年第5期40-46,共7页
Modern Electronics Technique
基金
国家自然科学基金资助项目:空地机器人网络的同时视觉目标定位与分布式运动规划(61503118)
国家自然科学基金资助项目:基于脑认知与机器感知联合决策的脑-机融合性能优化研究(62006135)。
关键词
多目标跟踪
改进DeepSort
轨迹预测
外观特征生成
图像处理
对比实验
multi-target tracking
improved DeepSort
trajectory prediction
appearance feature generation
image processing
contrast experiment